n-gram hashing for speculative decoding

[nightscape]

Hi, I hope I didn't miss an existing issue.
Does MLX already have a generic n-gram hasher for speculative decoding like ggml-org/llama.cpp#19164 ?
I saw #433 and #833, but they seem to be LongCat specific.
IIUC, the llama.cpp n-gram hashing works for arbitrary models and seems to bring quite a speed up in cases where parts of the response are verbatim copies of previous responses (which can happen quite a lot when iteratively refining code).

[nightscape]

@awni does this exist already?
If not, would it be worthy of a PR?

[TomLucidor]

Someone will need to dip their hands in them first

[mayank2130]

hey, i'm working on adding PLD to MLX-LM.

Instead of a draft model, use n-gram lookup over token history to propose draft tokens — same verifier path, just a different draft source.

Plan:

• ngram-simple: search recent n-gram window in token history, copy following tokens as draft
• Adaptive gate: skip speculation on prompts unlikely to benefit (low repetition density)
• CLI: --draft-type {none,ngram-simple} / server: draft_type in request body
• Existing speculative_generate_step verifier stays untouched

Open to discuss design.